CuTeDSL operators dependencies

[StableLlama]

Running the code with its standard dependencies installed I run into an error telling me that nvidia-cutlass-dsl and apache-tvm-ffi are missing.

Installing them with

pip install --no-build-isolation --no-cache-dir nvidia-cutlass-dsl apache-tvm-ffi 

the code is running, but showing in the log:

cutedsl version 4.7.1 is not known-good (ok: {<Version('4.4.2')>, <Version('4.4.1')>, <Version('4.5.2')>}); set TORCH_NATIVE_SKIP_VERSION_CHECK=1 to override

I guess that there should be either the dependencies of ST updated and/or the documentation updated and also include a hint about good versions.

[bghira]

it's a version conflict upstream that is not easily solved:

• PyTorch native operators declare TVM-FFI as optional, without them it just disables those optimized overrides and fall back to ATen. the performance impact of that's unknown and i dunno what operations want it enabled, so, empirically testing that is difficult.
• TorchAO MXFP8 MoE declares TVM-FFI as required and can hard fail, but i don't think we're hitting that path here, and it's easily disabled by switching from torchao to eg. sdnq which relies on triton
• transformers SonicMoE declares TVM-FFI as required and requires nvidia-cutlass-dsl>=4.6.0 and apache-tvm-ffi>=0.1.10.

so the conflict is:

• PyTorch 2.13 recognizes CuTeDSL 4.4.1, 4.4.2, or 4.5.2
• transformers SonicMoE requires >=4.6.0

installing cutedsl 4.7.1 satisfies transformers, but makes pytorch emit the warning you've seen.

if it's crashing, give a traceback of where the issue is getting raised from. but if it's just printing to the logs, the question of whether to fix it is a bit harder to answer, and it depends on whether the performance improvement is more than the burden of fixing and testing it.

[StableLlama]

I only have that warning. When I can just ignore it, I'm happy with that.